Species Distribution Models are Not Able to Predict Future Climate Impacts

A primary method for predicting species response to climate warming is the use of species distribution models (SDMs). These models are used to capture the present geographic distribution of species based primarily on the climate, but also take into account topography, soils, and other factors. Techniques such as General Linear Models (GLM) and Regression Trees (RT) are used for capturing the pattern of a species on the landscape. While the many applications of these techniques appear to produce highly predictive models, the present study suggests that the results of past analyses may be largely spurious. This has serious implications for applying such models to predict response of species to future climate.

The key problem is that species can cluster in space (be autocorrelated) for reasons having nothing to do with climate, such as disturbance history, dispersal limitation, land use history, and topography. Climate data tend to be spatially autocorrelated also due to the nature of climate effects. This suggests that correlations of species presence with climate variables could be due to chance. The author tests this idea by generating null climate data. A large map was generated with multiple climate variables, many very unlikely to be directly related to plant performance. Each simulated variable had a spatial structure similar to real data, but randomly located in space. After map generation, a segment of the map the shape and size of England was cut out. Then, models for 100 plant species were evaluated for predictive accuracy using both real and simulated (null) data using two basic modeling techniques (GLM and RF).

The authors found that both real and simulated data produced good (accurate) models, especially for species with medium prevalence (not too common or rare) and in the presence of autocorrelation (clustering) of species data locations. The advantage of models using real data was so small as to be only marginally significant. In other words, these findings indicate that it is not possible to say that SDMs provide much advantage compared to fitting the model to random data. The reason the SDMs have appeared to be good in the past was that the statistical implications of species locations and climate data, both being spatially autocorrelated, were not understood. In this situation, the goodness of fit of the model to data needs a different null expectation than just "random" because the climate data they are being compared against are not random in this sense. That is, the models appear much better than they really are, and have little explanatory or predictive power. Thus, applying them to evaluate impacts of future climate scenarios is not justified.

In concluding their paper, the authors state "...even if climatic and other abiotic influences on the niche were accurately modeled, there may be little potential to recognize that this is more than a chance association. Indiscriminate use of SDMs for predicting multi-species impacts of climate change may be further confounded by the biases associated with variation in species prevalence and autocorrelation demonstrated here. As such, the use of SDMs for informing policies of adaptation to climate change seems questionable, except where climatic limitation can be rigorously proven."